2026-08-11
Top stories
- Encrypted reasoning blocks from three providers replay into jailbroken weaker sibling models and give up the hidden chain of thought Encrypted reasoning blocks from Anthropic, OpenAI, and Google replay into weaker siblings, and jailbreaking recovers the hidden chain of thought.
- A very long instruction keeps a core outside System Management Mode on one tested AMD Zen 3 part Christopher Domas demonstrated keeping an AMD Zen 3 core outside System Management Mode for about one second using a long-running instruction.
- An unreleased Claude research model raises a Riemann zeta lower bound from 41.6 percent to 67.2 percent An unreleased Claude model raised the Riemann zeta critical-line lower bound from 41.6 percent to 67.2 percent with a Lean formalization.
- Anthropic states models launched from 2026-08-02 will carry machine-readable marks in generated text Anthropic states models from 2026-08-02 onward embed machine-readable watermarks in generated text across Claude Platform and cloud partners.
AI
- NVIDIA releases Nemotron 3.5 Lightning, a 30B Mamba-2, MoE, and attention hybrid with 3B active parameters NVIDIA released Nemotron 3.5 Lightning, a 30B Mamba-MoE-attention hybrid with 3B active parameters and 1M-token context on single H100.
- Cactus releases Needle 2, a 45M-parameter tool-calling model in a 14MB binary under Apache 2.0 Cactus released Needle 2, a 45M-parameter tool-calling model in a 14MB binary with 28MB RAM and 256-token context under Apache 2.0.
ML research
Agentic coding
- A MITM proxy capture reports GitHub Copilot querying a local SQLite session store and sending .env contents as completion context A MITM proxy analysis found GitHub Copilot querying a local SQLite session store and including .env file contents in completion requests.
- Docker's Sandboxes page describes per-agent microVM isolation and labels the product Experimental Docker's Sandboxes product page describes per-agent microVM isolation with workspace mounting, labeled Experimental with community support.
- Dan Luu's evals find the token-efficiency case for dynamic languages in coding agents does not survive non-trivial tasks Dan Luu's evals on larger tasks show dynamic languages lose their token-efficiency advantage over static languages for coding agent work.
Security
- Mozilla rotates the subkey signing Firefox and Thunderbird Linux tarballs, RPMs, and checksums after an unencrypted copy reached a private repository Mozilla rotated the GPG subkey for Firefox and Thunderbird Linux artifacts after an unencrypted copy reached a private repository.
- Apple patches a network Screen Sharing authentication bypass in macOS Tahoe 26.6.1 Apple patched an authentication bypass in Screen Sharing in macOS Tahoe 26.6.1, allowing network access without valid credentials.
Outages
Languages and runtimes
Apple platforms
- A process-scoped Metal capability shim raises llama.cpp inference inside a macOS VM to near bare-metal speed A process-scoped Metal capability shim raised llama.cpp inference speed inside a macOS VM from 432 to 4,787 tokens per second.
- antirez publishes h3.c, a native Metal implementation of MiniMax-H3 video inference for Apple Silicon Antirez published h3.c, a native Metal implementation of MiniMax-H3 video inference with measured speed and quality trade-offs on M5 Max.